Google DeepMind Frontier Safety Framework

This document is a public update to Google DeepMind's Frontier Safety Framework (FSF), a voluntary internal governance protocol governing the evaluation and mitigation of severe risks from frontier AI models, with no stated legal basis beyond DeepMind's own AI Principles. The framework establishes Critical Capability Levels (CCLs) as the operational threshold concept, and the updated terms introduce tiered Security Level recommendations mapped to each CCL, a revised deployment mitigations procedure requiring a corporate-governance-reviewed safety case before general availability deployment, and a new approach to detecting and mitigating deceptive alignment risk through automated monitoring. The document is a voluntary corporate governance commitment rather than a legally binding agreement with users or third parties, and its provisions reflect internal operational standards rather than enforceable contractual obligations; the framework's practical force depends on DeepMind's internal governance structures and is not independently verifiable from this document alone. The document references the Seoul Frontier AI Safety Commitments as an external reference point and states that DeepMind aims to share information with government authorities when a model reaches an unmitigated CCL posing material public safety risk, engaging discussions relevant to the EU AI Act's high-risk AI classification and systemic risk provisions, as well as broader national AI governance frameworks in the UK and US; the document does not specify which jurisdictions or regulatory bodies would receive such disclosures or under what conditions.